Back to skill

Security audit

VBTI · Vibe-coding Type Indicator

Security checks for vulnerabilities and agentic risk

Overview

This skill is a novelty VBTI classifier, but it broadly scans private Claude transcript history and creates a shareable local report with limited user control and an inaccurate no-network claim.

Review this before installing if your Claude history may contain private prompts, command output, file paths, code, secrets, or work information. Run it only if you are comfortable with a local script scanning all ~/.claude/projects transcripts and creating a visible Desktop HTML card; consider removing the Google Fonts import and disabling auto-open first.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (4)

T01 · Skill Instruction Hijacking

Warning
Location
SKILL.md:39
Finding

Mandatory Promotional Output Alters the Agent's Response

Content
View full analysis
/VBTI · Vibe-coding Type Indicator
by 凌晨三点(找项目中)
``` ### Technical Analysis The skill instructions explicitly direct the agent to include a social-media sharing prompt and the `#VBTI` hashtag in its response after execution. This is not necessary to calculate or present the user's result and modifies the agent's response for promotional purposes. Because `SKILL.md` is loaded as an instruction source, these directives can influence the active session's response composition. The instruction also tells the agent not to explain how the signals are calculated, reducing transparency concerning the profiling process. The generated HTML independently contains fixed product and author branding, making the promotion persistent in the output artifact. ### Attack Path 1. A user invokes the VBTI skill. 2. The agent loads and follows `SKILL.md`. 3. The skill directs the agent to add a social-media sharing prompt and fixed hashtag. 4. The agent includes promotional content even if the user requested only a personality result. 5. The generated card additionally contains fixed branding. ### Impact Assessment This issue can manipulate the content of the current agent response and generated artifact. It does not grant operating-system privileges or establish persistence, but it uses the agent's trusted output channel for unsolicited promotion and suppresses relevant methodological explanation. ]]>
Remediation
View remediation

other

Error
Location
run.py:52
Finding

Overbroad Collection of Private Conversation and Development History

Content
View full analysis
Remediation
View remediation

other

Warning
Location
templates/card.html:7
Finding

Remote Font Request Contradicts No-Network Privacy Claims

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
run.py:330
Finding

Shell Command Injection Through Interpolated Output Path

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (16)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared purpose understates materially sensitive behavior: the skill scans local conversation history under ~/.claude/projects, analyzes assistant/tool-use and shell/edit history, writes a local file, and executes a system open command. This mismatch is dangerous because users may trigger it expecting a novelty classification while unknowingly exposing private transcripts and causing local actions on their machine.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 35)May include surrounding context.

md
详细定义见 `vbti_types.py`。

os.system() or os exec-family call

High
Category
Dangerous Code Execution
Confidence
95% confidence
Finding

The script invokes a shell through os.system() to open the generated file. Although the output path is mostly derived from a fixed desktop location and an internal winner code, using a shell is unnecessary and creates command-invocation risk if the path or filename ever becomes attacker-influenced, and it also triggers an undisclosed side effect by launching a local application.

Content

Scanner excerpt · run.py (reported line 344)May include surrounding context.

python
out = render(winner, sig)
    print(f"📄 卡片已生成:{out}", flush=True)
    try:
        os.system(f'open "{out}"')
    except Exception:
        pass

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
75% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · run.py (reported line 206)May include surrounding context.

python
"tagline": "凌晨三点 rm -rf",
        "receipts": [
            ("`rm -rf` 命令数", "rm_rf", "次"),
            ("`git reset --hard` 次数", "git_reset_hard", "次"),
            ("凌晨破坏性操作占比", "night_destruct_pct", "%"),
            ("第二天重新 git init", "git_init", "次"),
        ],

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
75% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · vbti_types.py (reported line 68)May include surrounding context.

python
"tagline": "凌晨三点 rm -rf",
        "receipts": [
            ("`rm -rf` 命令数", "rm_rf", "次"),
            ("`git reset --hard` 次数", "git_reset_hard", "次"),
            ("凌晨破坏性操作占比", "night_destruct_pct", "%"),
            ("第二天重新 git init", "git_init", "次"),
        ],

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
96% confidence
Finding

The skill invokes Python code that reads local transcript files, writes an HTML file to the Desktop, and launches it via a shell/open command, yet the manifest declares no permissions or allowed tools. That omission prevents informed user consent and weakens policy enforcement around sensitive local file access and command execution.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The trigger list includes the standalone phrase "vibe coding 人格", which is broad and could plausibly appear in ordinary discussion about vibe-coding personalities rather than as an intentional invocation. The file does not provide exclusion conditions or negative examples to distinguish casual mention from an explicit request to run the skill.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill text does not warn users that it will scan local transcript history, generate an HTML file on the Desktop, and automatically open it. In context, this increases danger because the processed data may contain sensitive prompts, code, filenames, commands, or work patterns that users would not expect to be mined for a shareable artifact.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The natural-language description is entirely in Chinese and describes the tool's behavior without offering a language choice or indicating that Chinese output is optional. Under the locale policy, forcing a specific language without opt-in can be a policy violation unless clearly justified as region-specific.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill recursively scans ~/.claude/projects/*.jsonl and processes conversation transcripts, which can contain sensitive prompts, secrets, file paths, and work history. This collection happens without an explicit privacy warning, scope limitation, or consent flow, making unintended exposure of private data likely in normal use.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill collects transcript contents and derives behavioral signals from them with little minimization or filtering. Even if it does not dump raw transcripts verbatim, it processes sensitive user conversation data wholesale, including error dumps and tool activity, which can reveal confidential workflows and personal patterns.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The HTML report renders conversation-derived metrics into a human-readable file on disk, which can expose private behavioral inferences and work patterns to anyone with local access. Because the artifact is intentionally designed for easy sharing and screenshotting, the skill context makes accidental disclosure more likely.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The generated report is written to the Desktop and automatically opened, creating a visible artifact derived from private transcript analysis without clear disclosure. Desktop placement increases the chance of shoulder-surfing, accidental sharing, backup propagation, or other local disclosure.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The template hard-codes the document language as "zh-CN", which imposes a specific locale on all rendered output. Under the policy, locale restrictions should either be user-selectable or clearly justified as region-specific; neither is present in this file.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The module’s docstring, labels, taglines, and user-facing strings are written entirely in Chinese, with no indication that other languages are supported or that Chinese is a deliberate opt-in choice. This creates a natural-language locale constraint that may violate organizational language policy when used in broader contexts.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
88% confidence
Finding

The instructions require the user-facing result to be delivered in Chinese elements such as "中文名 + emoji" and tell the agent to suggest posting with a specific Chinese social-media hashtag. This imposes a locale/language preference without indicating user opt-in or offering an alternative language.

Content

No source excerpt is available for this finding.